Neural networks that fire spikes

“Spiking neural networks” is a bit of a mouthful. Can you break it down for us?


In the brain, the individual neurons are connected to form neural networks. Our work imitates this approach to produce artificial intelligence (AI). Most of the work to date has been with deep neural networks, in which the transmission of signals between neurons is either analog – at a specific voltage, for instance – or digital via discrete packages of information. Spiking neural networks function in much the same way, only they have a different way of transmitting information: as the name suggests, the individual neurons in these networks emit – or “fire” – nerve impulses – known as “spikes” – at one another. This puts us much closer to what happens within the brain. The information resides in the spike rate; for instance, a high signal value is represented by multiple spikes fired in quick succession. We’re currently researching the most efficient way of doing this as part of a large-scale project involving a further ten Fraunhofer Institutes.


So for spiking neural networks, you’re effectively charting a middle course between analog and digital. Why is that?


Analog and digital signal processing each have their advantages. Analog processing is very energy-efficient, while digital transmission offers better error correction and good scalability. Spiking neural networks allow us to combine the best of both worlds. We can pair analog signal processing with binary transmission, which opens up a wealth of possibilities. According to our research hypothesis, this will allow us to produce very energy-efficient circuits that are ideal for Edge AI applications in which the “intelligence” is transferred to the devices themselves. Energy efficiency is a key requirement. 


What are spiking neural networks particularly good at?


If you’re using a computer to evaluate large image data sets, deep neural networks are currently still the best option. Spiking neural networks really come into their own when implementing hardware. In other words, wherever neuromorphic hardware is called for – special, extremely energy-efficient chips that pave the way for “intelligent” devices. Depending on the application, it may be that analog, digital or spiking neural networks offer the best performance. SNNs are particularly suitable for time series analysis. We almost always encounter such time series whenever sensor data needs to be evaluated – for instance in system monitoring in manufacturing, medical ECG tests or even audio signals. So SNNs are another tool in our toolbox. 


Where are spiking neural networks used?


At the moment, SNNs are being tested most commonly as part of research projects, in predevelopment and by start-ups. The future will tell where they find their niche.